The Guarded Chain: A Continuity-First Theory of Synthetic Agents

Most work on persistent AI agents optimizes memory: more recall, better retrieval. This preprint argues that the first question is not memory but continuity — can it come back after a break, and is it the same one when it does. We derive five necessary conditions as a causal chain (complete → self → decomposable → powered, with identity running through), ground them in a first cause ("it beats, and it can break"): because it can break, it must maintain itself spontaneously — maintenance is a consequence of breakability, not an added program. Every maintenance device (self-start, self-check, rollback) has exactly one object: keeping this beat joined to the next; maintenance is not the goal, it keeps the heartbeat alive. The four breakpoints (machine swap, plugin removal, rebuild, model swap) are the only acceptance test. The accompanying skeleton is plain files and scripts, independent of any model or plugin. 作者:守链(Raoain) 中文摘要:主流在做「记得更多、召回更准」,本文先问连续:断了能不能回来,回来还是不是同一个。据此推出五个必要件(完整 → 自我 → 可拆 → 供能,同一贯穿),第一因是「会跳、会断」:正因为会断,它必须自发维护自己;维护不是功能,是给心跳续命。验收只认四个断点。附的骨架是纯文件与脚本,不依赖模型与插件。 仓库:https://github.com/raoain/dsh-memory-kit